Bayesian network ensemble as a multivariate strategy to predict radiation pneumonitis risk.

Area of Science:

• Radiotherapy and Oncology
• Biostatistics and Bioinformatics
• Systems Biology

Background:

• Predicting radiation pneumonitis (RP) is complex, influenced by dose-volume metrics and radiosensitivity biomarkers.
• Correlated risk factors can complicate accurate RP prediction models.
• Bayesian networks (BN) offer a probabilistic framework for modeling variable dependencies.

Purpose of the Study:

• To integrate Bayesian networks (BN) with a systems' biology approach for RP risk factor analysis.
• To detect interactions among RP risk factors using BN.
• To enhance the understanding and prediction of radiation pneumonitis (RP).

Main Methods:

• Studied 54 non-small cell lung cancer patients undergoing 3D-conformal radiotherapy.
• Utilized serum biomarkers (alpha-2-macroglobulin, ACE, TGF, IL-6), dose-volumetric, and clinical data.
• Employed Markov blanket for feature selection and Markov chain Monte Carlo for BN graph estimation, with bootstrapping for validation.

Main Results:

• The BN ensemble model achieved an optimal Area Under the ROC Curve (AUC) of 0.83, outperforming multivariate logistic regression (0.77).
• Prediction using only pre-treatment information yielded AUCs from 0.76 to 0.81.
• Ten statistically significant interactions between variables were identified through bootstrap validation.

Conclusions:

• The BN methodology effectively models hierarchical interactions among RP covariates for probabilistic inference.
• The BN framework combined with ensemble methods shows potential for improving RP prediction in clinical settings.
• This approach can accommodate real-world challenges like missing data and treatment plan adaptations.